Relational coherence

Things we call accidents usually involve several interacting quantities changing at once. The question I started with was whether an observer who understood the relationships between those quantities could compensate — not by abandoning what they were doing, but by adjusting the relationship and carrying the original action through.

Centrality here is informational, not geometric. Being central to a system means standing in a particular decision and information relationship to it, not standing in the middle of it.

The strong claim — that an observer who knew every parameter could prevent any accident — does not survive contact with the counterexamples. Knowing does not create the authority to act, information goes stale before an actuator responds, and two safety constraints can be arranged so that no action satisfies both. None of those are information problems.

Knowledge does not create corrective authority.

So the question had to be narrowed into something that could be answered: when does a safe policy exist, which relationships does it have to understand, and what physical reserves must still be available to carry it out?

A finite-horizon, partially observed, robust viability problem. The state is not only physical: it carries the relationship configuration, the remaining reserve, the commands still in flight, the admissible set of hidden parameters, and what the observer actually knows. The question becomes whether a policy exists that keeps that whole augmented state inside a safe region for the horizon.

That is a reformulation, not a discovery. The value of it is that it is falsifiable, and that it makes the conditions explicit rather than leaving them as an intuition about understanding.